[OpenCL:Bugfix] Use vload_half/vstore_half builtins in buffer_convert_buf

[Qxinyu]

Summary

• Fix issue Qwen2.5-Omni-3B/7B 在最新代码上opencl上推理崩溃 #4802: Qwen2.5-Omni crashes on OpenCL at buffer_convert_buf kernel build on devices whose driver advertises cl_khr_fp16 but rejects direct scalar access to __global half* ("loading directly from pointer to type 'const __global half' requires cl_khr_fp16. Use vector data load builtin functions instead").
• In buffer_convert_buf.cl, route all scalar/vector half element access through vload_half / vstore_half / vload_half4 / vstore_half4 builtins via new MNN_LOAD_INPUT / MNN_STORE_OUTPUT / MNN_LOAD_INPUT4 / MNN_STORE_OUTPUT4 / MNN_LOAD_FLOAT macros. Covers buffer_convert_to_buffer, buffer_copy_to_buffer (incl. PACK_LEAVE tail loop) and the conv/dw filter conversion kernels (20 direct FLOAT* dereferences). Non-half types keep the original direct/vload4 paths unchanged.
• C++ side defines -DINPUT_IS_HALF / -DOUTPUT_IS_HALF wherever half types are selected (AddBuildOptionOfDataType in BufferConvertor, tensor-derived types in OpenCLRuntime::buildKernelWithCache, which also covers Softmax/ArgMax callers of these kernels).
• Null-check every buildKernelWithCache result in BufferConvertor so a kernel build failure degrades to a logged error instead of a segmentation fault (the crash path in the issue).
• Bump the buffer_convert_buf md5 in opencl_source_map.hpp to invalidate stale compiled-kernel caches.

Test plan

• build_ocl full op suite on OpenCL backend: 248/248 passed (fp16 path active on Apple GPU)
• softmax / argmax / convolution OpenCL op tests passed (exercise buffer_convert_to_buffer with half NC4HW4 unpack/pack)
• Needs verification on the reporter's device (星睿 O6) with Qwen2.5-Omni-3B

[CLAassistant]

Thank you for your submission! We really appreciate it. Like many open source projects, we ask that you sign our Contributor License Agreement before we can accept your contribution.
_{You have signed the CLA already but the status is still pending? Let us recheck it.}